Ephemeris.
Frontier
The frontier moved on two fronts at once — a model that can code unattended for days, and the systems being built to live alongside models like it: the defenses, the orbital compute, the company knowledge, the autonomy.
Fable for all of us. Mythos for a few.
Anthropic shipped Fable 5 to general availability and, beside it, Mythos 5 — the same underlying model with the safeguards lifted, gated to vetted cybersecurity and (soon) biomedical teams. Anthropic's own line: its capabilities exceed any model they've made broadly available. Stripe ran a Ruby migration through it that a team had scoped at two months and finished in a day.
# claude-fable-5 · generally available 10 Jun pricing $10 / $50 per Mtok (< half of Mythos Preview) free until 22 Jun on Pro / Max / Team / Enterprise frontier-code #1 among frontier models @ medium effort long context 3× Opus 4.8 on memory-augmented tasks vision beat Pokémon FireRed with no helper harness mythos-5 cyber + bio access only · 30-day retention
Assume the patch arrives late.
Cloudflare's argument from running its own stack as customer zero: against attackers wielding frontier models, the architecture around a flaw decides more than how fast you patch it. The reframe is defense-in-depth as a budget — when one layer misses, the next limits what the attacker can see, reach, or change.
- Score by ML on each request — novel exploit variants flagged before a CVE exists
- API Shield validates against your API's definition instead of a blacklist
- So thousands of generated payload variants collapse to "not in the schema"
A datacenter with a 70-metre wingspan.
SpaceX put out a half-hour technical briefing on its AI1 satellites ahead of the IPO: each one carries roughly the compute of 72 GB200s, powered by a solar array the length of a basketball court, built to clip together into larger constellations. The pitch is training and inference that runs where the power and cooling are free — orbit.
The bottleneck is context, not the model.
In a Stanford MS&E 435 lecture, Yash Patil of Applied Compute makes the case that the hard part of enterprise AI is not the base model — it's wiring in the knowledge a company already holds. Specialised, domain-tuned models routinely beat a stronger generic one once that internal context is in reach.
"A weaker model that can see your data beats a stronger one that can't."
The work that moves the needle sits upstream of the prompt: extracting tacit process knowledge, resolving which of three conflicting internal docs is current, and feeding that into retrieval an agent can trust.
The takeaway for a founder or staff engineer is to spend the next quarter on the context pipeline — sources, freshness, access — before reaching for a bigger model.
Cargo that flies itself.
Airbus showed a prototype pilotless cargo helicopter, with a first flight demonstration slated before year-end. The interesting part is not the airframe but the removal of the seat: once the route and the cargo are the whole job, autonomy stops being a driver-assist feature and becomes the operator.
That's today.
A lighter issue — the feeds turned over slowly overnight, so today is five picks with real signal rather than ten with filler.